Keep your AI setup in a folder you own

What my setup actually looks like

I use AI for things like Reel scripts, websites and video editing. Those jobs need different context. A Reel needs my writing voice. A website has its own files and design decisions. I keep that work in separate folders.

Here is a shortened view of my actual Reel’s Script folder. These are real file names. I have left out the research archive and other supporting files so you can see the useful part.

Reel's Script/
  AGENTS.md
  knowledge/
    VOICE.md
    PLAYBOOK.md
  drafts/

AGENTS.md tells the assistant how to work in that project. VOICE.md describes how my writing should sound. PLAYBOOK.md holds the writing principles. drafts/ holds the work. For example, my voice notes explicitly rule out fake personal tests and made-up time savings.

The download is a smaller starter version of this approach, made for this guide. It is not a copy of my entire private workspace. You get a short example of my real writing preferences and blank files for your own information.

1. Put the starter folder on your computer

Start with one real task: a welcome email, a Reel script, a proposal or a study plan. You do not need to organize your whole life before this becomes useful.

1. Download the starter ZIP. On Mac, double-click it to unzip. On Windows, right-click it and choose Extract All. Open the extracted folder, not the ZIP preview.
2. Move AI-Workspace into Documents, or another place you can find again. Keep this as your main copy.
3. Open START-HERE.txt. Then open 01-ABOUT-ME.txt in TextEdit on Mac or Notepad on Windows. Fill in the brackets and save. Repeat for the other two numbered files.

AI-Workspace/
  START-HERE.txt
  01-ABOUT-ME.txt
  02-HOW-TO-HELP.txt
  03-WRITING-STYLE.txt
  Projects/
    My-first-project/
      BRIEF.txt
      PROGRESS.txt
      References/
      Outputs/
  Prompts/
  Examples/

These are ordinary text files. No coding, terminal, extensions or automation setup required. My own files often end in .md, which means Markdown: plain text with simple formatting. The starter uses .txt so you can open it with an app you already have.

For your first run, leave My-first-project named as it is so the included prompts work unchanged. Later, duplicate that project folder, give it a useful name, and change the project name in your prompts too.

2. Give the AI something specific to work with

Name: Sam
I run a small coaching business.
My audience: people new to meal planning.
Priority: make my welcome emails clearer.

Use plain language.
Explain unfamiliar terms.
Ask when an important detail is missing.
Never invent results or testimonials.

Tone: friendly and direct.
Example: “Pick one dinner you already like.
Let’s make that your starting point.”
Why I like it: simple, useful, no big promises.

Make: a welcome email, about 150 words.
For: readers of my meal-planning checklist.
Use: the checklist in References/.
Avoid: a sales pitch or weight-loss promises.

Done: the first welcome email draft.
Approved: friendly tone, no sales pitch.
Latest: Outputs/welcome-email-v1.txt
Next: choose between the two subject lines.

Your background: 01-ABOUT-ME.txt

Write a few useful facts, not your life story. What do you do? Who is the work for? What are you trying to get done? Include your experience level so the AI knows what to explain.

Your preferences: 02-HOW-TO-HELP.txt

I included sensible starting rules, such as using plain language and not inventing results. Read them and change anything that does not fit you. A rule should describe a behavior you actually want.

Your voice: 03-WRITING-STYLE.txt

Paste one short piece of your own writing that you like. Add why you like it. For example: ‘This sounds like me because I get to the point and give an actual example.’ One useful example is a better start than twenty vague adjectives.

The job: Projects/My-first-project/BRIEF.txt

Keep the current task here so it does not get mixed up with your permanent preferences. Here is a fictional example to show the level of detail:

Project: Welcome email
Audience: People who downloaded my beginner meal-planning checklist
Deliverable: One email, about 150 words
Goal: Help them take the first step in the checklist
Facts: Use only the attached checklist
Tone: Friendly and practical
Avoid: Weight-loss promises, sales pitch, invented personal stories
Today's task: Draft the email and suggest two subject lines

Put the actual checklist in References/ before using a brief like that. Naming a document does not give the AI access to it.

3. Bring over the useful parts of an old chat

If an existing chat already knows your project, open it and paste this. It gives you a draft to review, not a guaranteed export of every memory the app holds.

02-RESCUE-OLD-CHAT.txt

Help me move the useful context from this conversation into my own files.
Use only information you can actually access. Do not claim to export hidden
memory, other chats or unavailable project instructions.
Draft separate plain-text sections for:
1. ABOUT ME: confirmed facts, audience and priorities.
2. HOW TO HELP: preferences and corrections I explicitly gave you.
3. WRITING STYLE: preferences plus short examples of my own writing from this chat.
4. PROJECT BRIEF: current goal, requirements and source references.
5. PROJECT PROGRESS: approved decisions, useful outputs, open questions and next step.
Distinguish my words from your suggestions. Do not turn an unapproved idea into
a decision. Mark anything uncertain for me to review. Remove secrets and private
third-party information. Show the drafts here before saving or replacing files.

1. Read the result. Remove invented details, old priorities and suggestions you never approved.
2. Copy the useful sections into the matching files. Check your app’s visible custom instructions and saved memories separately, if available. Copy only the parts you still want.
3. Save the latest useful draft in Outputs/. Add its exact filename and the next action to PROGRESS.txt.

You can skip this step if you are starting fresh. You do not need an export of your entire chat history. Start with the context needed for the job in front of you.

4. Let the AI read the files

An AI model is the engine. The app decides whether it can open files on your computer. Switching models inside one app and moving your work to a different app are different actions. This folder helps with the second one.

The simplest option: attach or paste

1. Open a fresh chat in the AI app you already use. If it supports file uploads, attach the three numbered text files plus BRIEF.txt and PROGRESS.txt from your project.
2. If uploads are unavailable, paste their contents into the message with each filename above its text. Add only the references needed for this task.
3. Paste the start prompt below. When the answer is useful, save it back into your local folder yourself.

An uploaded file is a copy. Editing your local file usually does not update that attachment. Upload the updated version again in the next chat. Pasting a folder path into an ordinary chat does not grant access to your computer.

If your desktop app supports local folders

In Codex or ChatGPT desktop with local project support, open the project menu, choose Edit project, then Add folder. Select AI-Workspace. Use Make primary if another folder is currently primary. Start a task in that local project.

In Claude Desktop with Cowork available, use Work in a folder and select AI-Workspace. Features and labels can vary by account and app version. If you do not see a folder option, use the attach-or-paste method above.

Connect only the workspace you need. Keeping your master files on your computer does not mean an online AI processes them offline. Relevant content can still be sent to the provider. Keep passwords, API keys and private client details out of this starter folder.

01-START.txt

I want to work on My-first-project.
Read 01-ABOUT-ME.txt, 02-HOW-TO-HELP.txt, 03-WRITING-STYLE.txt,
Projects/My-first-project/BRIEF.txt and PROGRESS.txt.
Use the files I attached if you do not have folder access.
Do not assume you can read a file just because I named it.
List the files you actually read. Summarize my audience, the deliverable,
two writing preferences and the next action. Flag blank fields or conflicts.
If a required file is unavailable, tell me exactly which one.
Do not begin the deliverable until we have checked this summary.

Check its summary against your files. Did it name your actual audience, deliverable and preferences? If it missed something, fix the attachment or folder access before asking for the full task. A confident ‘I read everything’ is not enough.

5. Finish one small task

Once the summary is right, give it the task from your brief. For the fictional email example, you could say:

The summary is correct. Read the checklist in References/ and draft the
welcome email from BRIEF.txt. Give me two subject lines. Use the style
example in 03-WRITING-STYLE.txt. If the checklist is missing, stop and tell me.

Review the result. If it is too formal, say which sentence sounds wrong and how you would say it. Save the useful draft in Outputs/welcome-email-v1.txt. If the AI can write files, ask it to save there and check that the file actually appeared. Otherwise copy and save it yourself.

For a correction you want to keep using, add one sentence to your writing-style file. ‘Use contractions and get to the useful part in the first paragraph’ is more actionable than ‘make it human.’ Do not save every one-off edit as a permanent rule.

6. Save where you got to

This is the part that makes the next session easier. Before you close a useful conversation, paste the save-progress prompt.

03-SAVE-PROGRESS.txt

Prepare a short handoff for the next session using this conversation.
Include the date, current goal, what we finished, decisions I approved,
the exact names of useful files, unresolved questions and one next action.
Do not include rejected drafts, secrets or guesses presented as facts.
Show proposed changes to PROGRESS.txt first. Suggest changes to my profile or
writing rules separately; do not save one-off choices as permanent preferences.
After I approve, update Projects/My-first-project/PROGRESS.txt if you have file
access. Otherwise give me the exact text to copy there and say it was not saved.

Check the handoff, then save it. PROGRESS.txt should say what is settled and what comes next. Your next AI does not need to wade through five rejected email drafts to learn which version you approved.

Updated: [today's date]
Goal: Send a helpful welcome email to checklist readers
Finished: First email draft
Approved: Friendly tone; no sales pitch
Latest output: Outputs/welcome-email-v1.txt
Open question: Which subject line to use
Next action: Compare the two subject lines and choose one

That is a fictional example of a handoff. Replace it with what actually happened in your project. Keep older outputs if you want to compare them, but make the current filename unambiguous.

7. Try continuing in a fresh chat

Save your current files first. Open a fresh chat, either in the same app or a different AI app. Connect the same folder, or attach the updated files and the latest output. Then paste this:

04-SWITCH-AI.txt

We are continuing an existing project in a fresh conversation.
Read the attached files or the files in the folder I connected:
01-ABOUT-ME.txt, 02-HOW-TO-HELP.txt, 03-WRITING-STYLE.txt,
Projects/My-first-project/BRIEF.txt and PROGRESS.txt.
Then read the latest output named in PROGRESS.txt if it is available.
List what you actually read. Tell me the goal, approved decisions, open questions
and next action. Quote one short writing preference from my files so I can check
that you loaded the right version. Flag missing files and conflicts.
Do not restart the project or invent missing history. Wait for me to confirm
the summary before continuing.

1. Check that it names the correct goal and next action.
2. Check that the writing preference it quotes really exists in your file.
3. Check that it knows which decisions are approved and which questions remain open. Then ask it to continue.

You have moved useful context. You have not transferred every hidden memory, app integration, permission or conversation. Different models will still write and reason differently. The win is that the next one has a clear starting point.

When something feels off

It says it cannot see my folder

Use the app’s actual folder picker or attach the files. If you renamed My-first-project, update the names in your prompts. A local file path by itself is not an upload.

It still sounds nothing like me

Check that it read 03-WRITING-STYLE.txt. Replace vague preferences with a short example of your writing and one specific correction. Ask it to revise one paragraph before generating the whole piece again.

It remembers the wrong version

Check the date and contents of PROGRESS.txt. Remove contradictory old rules. In an upload-based chat, start fresh and attach the current files. Say which file is the latest.

It claims it saved a file, but I cannot find it

Look in the actual folder. Ask for the exact saved path. If file writing is unavailable, ask for the text and save it yourself. A message that says ‘saved’ is not the file.

My folder is getting huge

Keep the profile and progress notes short. Separate unrelated projects. Load the active brief, progress and relevant references first. Add an archive folder later if you need it; do not ask the AI to read everything on every task.

Do I need AGENTS.md, CLAUDE.md or skills?

You can start without them. Some apps recognize special instruction files, but their rules differ. The start prompt explicitly names the files so this kit does not depend on one app’s automatic behavior. Add app-specific helpers later if they make your actual workflow easier.

How do I avoid losing the setup?

Keep a second copy using your usual computer backup or an external drive. Back up before a big reorganization. If you use cloud sync, check that the files are available on your computer and remember that deletions can sync too.

Your first session is done when…

1. The three numbered files contain your own information and a real writing example.
2. BRIEF.txt describes one actual task, and the AI can accurately summarize it.
3. You have saved one useful output and an up-to-date PROGRESS.txt.
4. A fresh chat can identify the next action from those files without inventing missing history.

Start there. You can add more later, once you know what you keep needing. The useful part of my setup is having instructions, reference material and finished work that I can open and update myself.